LINUAN/fold_towels_recap_value

fold_towels_recap value network — DISTRIBUTIONAL (201-bin) head

Distributional value network for the SO-101 fold-towels RECAP (pi-star 0.6) reproduction. This is the 201-bin softmax (categorical) head, the winning A/B variant.

• Model class: RECAPValueNetwork (see recap_value_network.py)
• Config: RECAPValueNetworkConfig (num_value_bins=201)
• Backbone: PaliGemma gemma_2b, vision encoder + backbone frozen, 3 backbone layers unfrozen
• Data: LINUAN/fold_towels_recap_corpus_v2 (276 episodes, val split 0.1)
• Trained 3 epochs; this is the epoch-3 checkpoint (best val_mae):
  • val_mae = 0.05278 (epoch 1: 0.05414, epoch 2: 0.05291)
  • val_loss (CE) = 3.99175

A/B result: this 201-bin head beats the scalar (1-dim MSE) head (fold_towels_recap_value_scalar, val_mae 0.05776) by ~0.005 val_mae.

Load

from lerobot.rl.algorithms.recap_value_network import RECAPValueNetwork, RECAPValueNetworkConfig
import json
cfg = RECAPValueNetworkConfig(**json.load(open("config.json")))
model = RECAPValueNetwork(cfg)
from safetensors.torch import load_file
model.load_state_dict(load_file("model.safetensors"))